Parameters

The variables in an artificial neural network that are adjusted during training to minimize the error in the model's predictions.

How it works

Parameters are the learnable numerical values of a neural network — primarily the weights of each connection and the biases of each neuron. During training, the optimiser updates these values iteratively based on gradients computed by backpropagation. A model with more parameters can represent more complex functions, but also requires more data and compute to train effectively and more memory to store and run. Parameter count is commonly used as a rough proxy for model capability.

Why it matters

Parameter count has become the primary axis on which AI models are described and compared — '7B', '70B', '671B' parameters have become shorthand for capability tiers. Understanding what parameters represent clarifies why larger models are more capable (more representational capacity) but also more expensive (more memory and compute). Techniques like quantisation, LoRA, and MoE all exist to manage the tension between parameter count and practical deployment cost.